Enterprise AI now has a control plane. Databricks made Unity AI Gateway generally available on August 4, 2026, giving organizations a single governed layer over every AI asset they touch: agents, models, MCP servers, skills, coding assistants, and external tools. Customers including Rivian, Asana, and Edmunds are already live on the platform, and Databricks reports that over a quadrillion tokens have passed through the gateway in the past twelve months.
The launch matters for any team running more than one AI tool, because that is now virtually every enterprise.
Why Enterprise AI Needs a Separate Control Plane
The standard narrative frames enterprise AI as a model selection problem: pick the right frontier model, deploy it, and iterate. But the operational reality in 2026 is a governance problem. Three pressures compound each other.
Cost scaling is non-linear. Unlike seat-based SaaS, AI consumption pricing tracks tokens, and agent workflows can generate tokens at volumes that dwarf what procurement ever modeled. A retry loop gone wrong, a background agent churning on a large context, a coding assistant left on over the weekend: all of these hit the bill in ways that traditional cloud budget alerts do not catch in time.
Security perimeters are dissolving. Agents now carry credentials into sensitive systems, invoke MCP tools that create or modify records, and retain conversation traces that may contain PII or confidential data. The sandbox-escape incidents documented throughout early 2026, including the OpenAI long-horizon model research findings, made clear that model providers alone cannot enforce enterprise security boundaries. That enforcement has to happen at the infrastructure layer.
Model choice is accelerating beyond procurement cycles. A new frontier model or agent harness appears, on average, every month. Organizations that want to adopt the best tool for each job need a way to swap models without rebuilding governance from scratch. Right now, most teams either lock in to one provider (and accept the tradeoffs) or manage a patchwork of separate dashboards, separate audit logs, and separate budget controls.
Unity AI Gateway addresses all three with a single platform rather than three separate tools.
What Unity AI Gateway Does
The product is structured around three pillars that Databricks calls cost, control, and choice.
| Pillar | Problem Solved | Key Capabilities |
|---|---|---|
| Cost | Runaway AI spend with no unified view | Granular attribution by model, team, provider, and application; hard spend caps; proactive budget alerts; Smart Routing (Beta) |
| Control | Agents accessing sensitive data without unified audit trail | Runtime guardrails and contextual policies; Unity Catalog identity, permissions, and lineage; MCP server governance; agent trace protection |
| Choice | Vendor lock-in when new models arrive | Multi-provider model registry in Unity Catalog; open API compatibility; governance that extends across providers without rebuilding per-vendor |
The gateway sits between any AI workload and the models or tools it calls. Every request passes through Unity AI Gateway, which applies configured policies before the request reaches its destination and logs the interaction to Unity Catalog system tables. That log then feeds the cost dashboards, audit trail, and Genie-powered analysis that teams use to understand their AI estate.
Smart Routing: Spend Less Without Asking Developers to Think About It
The most practically impactful feature for teams managing mixed-model environments is Smart Routing, currently in Beta. Routing dynamically selects the model for each incoming request based on configurable factors: quality requirements, cost budget, latency targets, availability, and account-level spend position.
The effect is that a coding agent call that needs frontier reasoning gets Claude Opus 5 or GPT-5.6, while a classification or summarization call that does not gets routed to a smaller, cheaper model without the developer having to hard-code that logic. Organizations can set routing policies at the workspace, team, or use-case level.
Databricks customer Zepto, a rapid-delivery platform, described the impact directly: “Unity AI Gateway has handled our production scale of more than 100 billion tokens per month with no availability issues, while giving us visibility into AI spend and user-level controls. Rapid access to the latest frontier and open-source models, combined with easily queryable usage tables, helped us quickly evaluate new models and build custom analytics around AI usage.”
That combination, production-scale reliability plus the ability to evaluate new models without a separate governance project, is the core value proposition for teams running at volume.
Governing Data and AI in One Place
The second architectural choice worth noting is that Unity AI Gateway is not a standalone product. It extends Unity Catalog, the governance system that already controls data access, lineage, and permissions for Databricks customers. That means AI governance inherits the same identity model, the same permission system, and the same audit trail as data governance.
In practice: when an agent calls an MCP server to query a customer database, Unity Catalog already governs who can read that data. Unity AI Gateway adds runtime enforcement on the AI interaction layer, so the same policies apply whether access comes from a SQL query or an agent tool call. Enterprises do not need to configure two separate permission systems for the same underlying assets.
STRABAG SE, the European construction and engineering firm, put it this way: “Unity AI Gateway enables us to centrally govern LLMs at scale across our data products and coding agents. By providing a single governed access layer across model providers, along with comprehensive observability and budget controls, it makes it much easier to safely scale AI across our organization.”
This is the same architecture challenge that Snowflake addressed from its side with the Snowflake Cortex AI Gateway launched in late July 2026. Both products reflect the same insight: the governance problem for AI cannot be solved by model providers alone; it has to be solved at the data platform layer where enterprise context actually lives.
What This Means for Enterprise AI Builders
For teams building on Databricks, the GA of Unity AI Gateway removes the primary argument against broad internal AI access. The historical friction was not that employees could not use AI tools; it was that enabling widespread access created unacceptable exposure on cost, security, and data governance simultaneously. Unity AI Gateway addresses all three without requiring IT to choose between control and developer velocity.
For teams not on Databricks, the product is a signal about where the market is heading. Multi-model orchestration is becoming a platform feature, not a custom integration project. The organizations best positioned for the next phase of enterprise AI, as Enera has argued in our analysis of the enterprise AI implementation gap, are those that build governance infrastructure now, before the model count and agent count scale beyond manual management.
A few open questions remain. Smart Routing is Beta, not GA, which means the cost optimization case still requires account team engagement rather than self-serve configuration. The governance model is tightly coupled to Databricks and Unity Catalog, which is a dependency, not just an advantage, for teams running heterogeneous stacks. And MCP governance, while supported, reflects a protocol that is still evolving across the industry.
None of these are blockers for teams already on Databricks. For those not yet on the platform, Unity AI Gateway is the most complete argument the company has made yet for why enterprise AI governance belongs in the data warehouse, not the AI provider portal.
Unity AI Gateway is available now on AWS, Azure, and GCP. Documentation is live in the Databricks product site, and Databricks is hosting a governance webinar on August 13 hosted by CTO Matei Zaharia.